Google DeepMind releases EmbeddingGemma 2 open multimodal embedding model

DeepMind Research News, blog.google, and MarkTechPost reported that Google DeepMind has launched EmbeddingGemma 2, an open multimodal embedding model. The new model natively maps combinations of text, images, audio, and video into a unified embedding space. MarkTechPost specified that the model contains 740 million parameters and is built on Gemma 4. [1][2][5]

The model is tailored for privacy-first, on-device retrieval-augmented generation pipelines and search applications, according to DeepMind Research News and YourStory.com. The Decoder reported that Google claims EmbeddingGemma 2 outperforms rival embedding models that are twice its size. Additionally, MIXED Reality News reported that the release drops the Gemma license gate and gains 0.21 on multilingual text performance. [1][3][4][6]

The original text-only EmbeddingGemma model was introduced in September 2025 to help applications process information directly on consumer hardware, as FTMQ SI reported earlier. The initial release accumulated more than 20 million downloads, DeepMind Research News reported. FTMQ SI previously reported that EmbeddingGemma 2 maps five input modalities into a single 768-dimensional space, operates using 567MB of memory, and is available under an Apache 2.0 open-source license. [1][9]

The architecture utilizes encoder-decoder initialization, geometric embedding distillation from larger models, and checkpoint merging to maintain performance. It exhibits strong robustness to Q8_0 and Q4_0 quantization configurations as well as embedding truncation with minimal performance degradation. These characteristics make the architecture suitable for deployment in resource-constrained environments such as mobile phones, laptops, and desktops. [8]